Unmixing Based Hyperspectral Image Fusion V.R.S.Mani 1 , Dr.S.Arivazhagan 2 , S.Amutha 3 Associate Professor 1 , Principal 2 , PG Scholar 3 Department of ECE, National Engineering College, Kovilpatti 1, 3 , Mepco Schlenk Engineering College Sivakasi 2 vrsece@rediffmail.com 1 , amutha007.ece@gmail.com 3 Abstract— The spatial resolution of hyperspectral image is often low due to the imaging spectrometer and the spectral resolution of the panchromatic image is also low. If both hyperspectral and panchromatic images are fused together then both spatial and spectral resolutions can be enhanced. The proposed fusion method is based on a sparse projection based unmixing technique. This method has the superior balance between spectral preservation and spatial enhancement over some traditional fusion methods. In addition, the added sparse and NMF based unmixing model make the fusion more stable. This method first decomposes the hyperspectral image into endmember and abundance matrix. Then sharpens the abundance matrix using the panchromatic image. Finally it produces the high spatial and spectral resolution fused image. The performance of the proposed algorithm is evaluated using different performance measures Index Terms—Hyperspectral image, panchromatic image, endmember matrix, abundance matrix, matrix factorization. I. INTRODUCTION Hyperspectral Imaging sensors collect information about the imaged scene in approximately 200 spectral bands in the visible and infrared wavelength regions (400-2500nm). Due to its high spectral resolution hyperspectral data are useful for accurate detection and identification of minerals, vegetation and man-made materials. The hyperspectral data are also used in ground object classification, mineral exploration and identification of natural and man-made materials. On the other hand panchromatic image contains high spatial resolution but has insufficient spectral information. If both of them are fused together, an enhanced image with high spatial as well as spectral resolution can be obtained. Different methods were proposed [1]-[11] by researchers and they are discussed in Section 1. Proposed method is elaborated in Section 3. The results obtained using proposed method is discussed in Section 5. Conclusion and the future work are given in Section 6. II. RELATED WORK The simplest method used for fusing hyperspectral and panchromatic image is the arithmetic method, which is just the addition or multiplication of the original HSI (Hyperspectral image) and PI (Panchromatic Image). It takes less computational time, but the fused data has severe spectral distortion [1].Projection substitution based methods are the classical and most popular among the available methods. In this method HSI is transformed into some other space and then the transformed data obtained is replaced by PI. In this method also the resultant fused data obtained is distorted to some extent [2]- [4]. Spectral component substitution technique such as Intensity, Hue and Saturation (IHS) technique replaces the intensity component of the low spatial resolution image with PI. It’s widely used because of its fast computational ability. However the intensity band and panchromatic band often differ from each other to a certain extent and it results in color distortion [5], [6]. Spatial domain methods such as High Pass Filtering (HPF) could reduce the degree of spectral distortion compared to IHS. But HPF method transfers only the excess high spatial frequency components into all the spectral bands with low spatial resolution. It also causes spectral distortion, because the detailed information extracted from the panchromatic image differs from the information contained in the original HS image [7].Wavelet based techniques are also commonly used, but their performance depends mainly on the spectral resampling method used, which caused difficulty in improving the spatial resolution of all hyperspectral bands [8], [9].The spectral unmixing method like Coupled NMF is based on the unsupervised un mixing. Here low spatial resolution hyperspectral image and high spatial resolution multispectral image (MSI) are alternatively unmixed by NMF. By combining HSI endmember matrix with the MSI abundance matrix the fused image is formed. And this method also suffers from spectral distortion [10].The constraint NMF unmixing technique also generates abundance and end member of HSI. The abundance matrix of HSI if sharpened by PI. A constraint term which preserves the spectral information is added and the fusion problem is turned into a constraint optimization problem. Additionally the projected gradient algorithm is used to produce the optimum solution [11].But here the update procedure is not a stable one. To overcome all these limitations a sparsity based technique is proposed in the following Section-3. III. PROPOSED METHOD In order to obtain a fused high-spatial resolution HSI with a little spectral distortion, the fusion model must satisfy the following two conditions. The sharpening information extracted from PI must be injected into the original low spatial resolution HSI There should be some apparent spectral preservation. In the existing method PI is used as it is which contains the low frequency component that will lead to spectral distortion. To avoid this, in the proposed method the PAN image is first passed through a high pass filter to remove the low frequency components. After that the PAN image is fused with the International Journal of Applied Engineering Research ISSN 0973-4562 Volume 10, Number 9 (2015) © Research India Publications ::: http://www.ripublication.com 7224